近年來學童在學校遭受霸凌的事件層出不窮，遭受同儕霸凌尚且有老師可以制止，若是遭到教師霸凌，學童只能回家之後告訴家長，但憾事已經發生。根據兒童福利聯盟的調查和研究[1][2]，遭受霸凌的學童，會有情緒沮喪、感到自卑、沒有信心……等影響，甚至會有自殘、想要自殺等更嚴重的影響。因此要防止學童遭受霸凌，需要在霸凌當下立即阻止，透過主管或其他教師阻止惡劣教師的霸凌行為。
  AI技術正在蓬勃發展，本論文擬透過影像辨識、語音情緒辨識和自然語言處理的技術，完成一套即時霸凌監控的系統，來幫助學童免遭教師霸凌。這套系統的霸凌偵測，包括影像、語音及文字等三個通道的資料，透過人工智慧的技術來進行判斷。我們在教室裝設攝影機，攝影機拍攝的即時影像，透過影像辨識的技術判斷是否有肢體霸凌的行為；另外，我們亦利用麥克風接收老師的聲音，透過語音情緒辨識老師是否有嘲諷等霸凌的語氣。此外，並將語音轉成文字，再利用自然語言處理的技術判斷老師是否有言語霸凌的行為，若有肢體霸凌或言語霸凌的行為，則立即通知主管。
  其中影像辨識用到的技術有Google開發的Posenet，利用Posenet抓取人體關節點再透過我們設計的演算法，判斷人體的動作；語音情緒辨識利用Librosa提取聲音特徵，再接上LSTM當作語音情緒的分類器，來判斷語音是否有霸凌的情形出現；自然語言處理的部分則是利用Google的預訓練模型BERT，加上我們向幼兒園老師取得的霸凌語句和非霸凌語句訓練，利用BERT的下游任務接上簡單的分類器，使得BERT有能力判斷老師是否有說出語言霸凌的語句。

In recent years, there have been an endless stream of bullying incidents against schoolchildren in schools. There are still teachers who can stop bullying. If they are bullied by teachers, schoolchildren can only go home and tell their parents, but a pity has happened. According to the investigation of the Child Welfare Alliance and research [1] [2], schoolchildren who suffer from bullying will be affected by depression, low self-esteem, lack of confidence. To prevent schoolchildren from being bullied, it is important to stop bullying immediately and stop the bullying behavior made by teachers.
  AI technology is developing vigorously. This paper intends to develop a real-time bullying monitoring system through the technologies of image recognition, voice emotion recognition and natural language processing, aiming to detect the teacher bullying and send the alert message to the corresponding users, including kinder garden manager and parents. A camera is installed in the classroom. By analyzing the real-time image taken by the camera, the developed system can identify whether or not there is physical bullying through image recognition technology. The system also uses the microphone to receive the teacher's voice, identify whether or not the teacher has verbal bullying through voice emotion recognition and natural language processing technology, and notify the supervisor immediately if there is physical bullying or verbal bullying.
  One of the technologies used in image recognition is Posenet developed by Google. Posenet is used to capture human joint points, and then judge human actions through our algorithm. The developed system also adopts the speech emotion recognition which uses Librosa to extract sound features, and then applies the LSTM as the classifier of speech emotion. The part of natural language processing is to use Google's pre training model Bert, plus the bullying sentence and non bullying sentence training we obtained from kindergarten teachers. The Bert's downstream task is connected with a simple classifier, so that Bert has the ability to judge whether the teacher speaks language bullying sentences.

目錄
目錄	V
圖目錄	VII
表目錄	VIII
第一章、簡介	1
第二章、相關研究	3
第三章、背景知識	5
3-1技術介紹	5
3-1-1語音分析	5
3-1-2文字分析	6
3-1-3影像辨識	7
第四章、系統架構	9
4-1問題描述與目標	9
4-1-1解決的問題	9
4-1-2目標	9
4-2系統架構	10
第五章、實驗分析	24
第六章、結論	27
參考文獻	28
附錄-英文論文	29
  
圖目錄
圖 1 Posenet關節點圖	8
圖 2霸凌偵測系統架構圖	10
圖 3預處理階段	11
圖 4轉換語音格式程式碼	11
圖 5語音轉文字程式碼	12
圖 6文字分析階段	13
圖 7 BERT訓練期	14
圖 8 BERT Token程式碼	14
圖 9 BERT下游分類器程式碼	15
圖 10 BERT訓練程式碼	15
圖 11 BERT使用期	16
圖 12語氣分析階段	16
圖 13 LSTM語音模型訓練期	17
圖 14聲音提取程式碼	18
圖 15 LSTM程式碼	18
圖 16 LSTM語音模型使用期	19
圖 17 影像辨識階段	20
圖 18 Posenet程式碼	20
圖 19座標標準化	21
圖 20 關節點演算法程式碼	22
圖 21 後續處理階段	23
圖 22未揮動手臂	26
圖 23揮動手臂	26

表目錄
表 1混淆矩陣表格	25
表 2文字分析模型結果	25
表 3語氣分析模型結果	25

[1]M. Boulton, M. Trueman and L. Murray, "Associations between peer victimization fear of future victimization and disrupted concentration on class work among junior school pupils", British Journal of Educational Psychology, vol. 78, pp. 473-489, 2008.
[2]D. Hawker and M. Boulton, "Twenty Years’ Research on Peer Victimization and Psychosocial Maladjustment: A Meta-analytic Review of Cross-sectional Studies", Journal of Child Psychology and Psychiatry, vol. 41, pp. 441-455, 2000.
[3]C. Wei, H. Zhang, L. Ye and F. Meng, "A school bullying detecting algorithm based on motion recognition and speech emotion recognition," 2020 International Conference on Intelligent Computing and Human-Computer Interaction (ICHCI), 2020, pp. 276-279.
[4]J. Yadav, D. Kumar and D. Chauhan, "Cyberbullying Detection using Pre-Trained BERT Model," 2020 International Conference on Electronics and Sustainable Communication Systems (ICESC), 2020, pp. 1096-1100.
[5]M. S. Likitha, S. R. R. Gupta, K. Hasitha and A. U. Raju, "Speech based human emotion recognition using MFCC," 2017 International Conference on Wireless Communications, Signal Processing and Networking (WiSPNET), 2017, pp. 2257-2260.
[6]J. J. Lasiman and D. Puji Lestari, "Speech Emotion Recognition for Indonesian Language Using Long Short-Term Memory," 2018 International Conference on Computer, Control, Informatics and its Applications (IC3INA), 2018, pp. 40-43.
[7]G. Papandreou et al., "Towards Accurate Multi-person Pose Estimation in the Wild," 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 3711-3719.
[8]Z. Cao, G. Hidalgo, T. Simon, S. -E. Wei and Y. Sheikh, "OpenPose: Realtime Multi-Person 2D Pose Estimation Using Part Affinity Fields," in IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 43, no. 1, pp. 172-186, 1 Jan. 2021.